Editorial illustration for Study Finds No One-Size-Fits-All Strategy for Multi-Agent Communication
Study Finds No One-Size-Fits-All Strategy for...
Forget finding the perfect way for AI agents to talk. There isn't one.
A new study tested five standard communication strategies across different agent networks. The conclusion is blunt. No single protocol works best in every situation.
The only consistent winner was a simple idea: preserve the specific information the next agent in line needs to do its job. Everything else is noise.
We analyze five common inter-agent communication strategies across two MAS topologies, finding that no fixed strategy is universally optimal. Instead, effective inter-agent messages consistently preserve action-centered information needed by downstream agents. Building on this, we propose the PACT (Protocolized Action-state Communication and Transmission), which treats inter-agent communication as a public state-update problem and projects each raw agent output into a compact action-state record before it enters shared history.
Across different MAS topologies, PACT consistently improves the performance-cost trade-off, achieving comparable or stronger task performance with substantially fewer tokens. The gains extend to production coding harnesses: PACT lifts OpenHands' resolve rate at -10% tokens-per-resolved, and is resolve-neutral on SWE-agent while halving input tokens.
So the researchers built PACT, a method that treats every message as a public update to a shared state. It filters an agent's raw output into a tight record of action and state. The results are concrete.
Across different network setups, PACT matched or beat task performance while using far fewer tokens. Real tests on coding agents proved it. OpenHands solved more problems with fewer tokens.
SWE-agent kept its success rate steady while cutting its input token count in half.
The lesson is not about a new universal language. It's about economy. Tell the next agent what it needs to know, and nothing more. That principle works everywhere, even if the specific words don't.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives xAI used Anthropic’s Claude via personal accounts after access revoked for months
From Archives Study examines temporal preference concepts in large language models
Common Questions Answered
Why is there no one-size-fits-all communication strategy for multi-agent networks?
According to the study, different agent network configurations have varying communication needs, and no single protocol consistently outperforms others across all situations. The research tested five standard communication strategies and found that effectiveness depends on the specific context and requirements of each network setup rather than following a universal approach.
What is the key principle that PACT uses for multi-agent communication?
PACT treats every message as a public update to a shared state and filters an agent's raw output into a tight record of action and state. This approach focuses on preserving only the specific information that the next agent in line needs to complete its task, eliminating unnecessary noise from communications.
How did PACT improve performance in real-world coding agent tests?
Real tests demonstrated concrete improvements: OpenHands solved more problems with fewer tokens when using PACT, and SWE-agent maintained its success rate while cutting its input token count in half. These results show that PACT matched or beat task performance across different network setups while significantly reducing token usage.
What was the most consistent finding across the five communication strategies tested in the study?
The only consistent winner across all tested scenarios was the simple principle of preserving the specific information that the next agent in line needs to do its job. Everything else in the communication was identified as noise that could be filtered out without impacting performance.
